Cumulative return
All four models and full-investment SPY
Portfolio performance
Exact daily overlapping-sleeve accounting. Use the cost selector to compare zero-cost sensitivity, declared implementation costs, and stress cases.
SPY always visible
Drawdown
Focused model versus SPY
Return distribution and execution
Daily return shape, turnover, trade counts, and cost sensitivity are kept separate from signal statistics.
Daily return distribution
Focused model versus SPY
Daily turnover
Twenty-session overlapping sleeves
Signal quality
Cross-sectional rank IC, top-40 precision, decile ordering, yearly behavior, and excess returns are evaluated before portfolio compounding.
Decision-level information coefficient
Same-date Spearman correlation
Mean return by score decile
Decile 10 contains the highest model scores
Horizon search and matched controls
The original 5-60 session screen is shown beside the later exact non-overlapping 40-126 session attribution. Point estimates are not presented as an optimal horizon.
Equal-budget horizon screen
Gross CAGR, implementation CAGR, and rank IC
Selection contribution
Frozen Ridge top 40 versus matched random and schedule-matched SPY
40 and 50 sessions show the clearest selection contribution versus sector-matched random portfolios.
Every tested attribution horizon trails schedule-matched SPY before transaction costs.
No horizon has passed the preregistered standard for a unique, benchmark-beating optimum.
Tail-risk A/B experiment
The fully invested Ridge-rank control is compared with the frozen downside treatment and SPY. Reduced tail exposure is not enough if winner rejection destroys returns.
Treatment inactive
Cumulative return
Control, tail treatment, and SPY
Risk and classification outcomes
Precision, recall, CVaR, drawdown, and rejected winners
Selected model specifications
Fold-level feature counts and selection stability. This section does not substitute selection frequency for SHAP.